Skip to content

MIGRATION_V2_TO_V3.md

Latest commit

 

History

History
121 lines (81 loc) · 3.38 KB

MIGRATION_V2_TO_V3.md

File metadata and controls

121 lines (81 loc) · 3.38 KB

Migrating from v2 to v3

react-native-fast-tflite v3 is a major rewrite that migrates from TurboModules to Nitro Modules. The public API remains similar, but there are several breaking changes.

1. Install react-native-nitro-modules

v3 requires react-native-nitro-modules as a peer dependency:

# npm
npm install react-native-fast-tflite react-native-nitro-modules

# yarn
yarn add react-native-fast-tflite react-native-nitro-modules

Then rebuild native projects:

cd ios && pod install && cd ..

2. Delegates: string -> array

- const model = await loadTensorflowModel(source)
+ const model = await loadTensorflowModel(source, [])

The delegate parameter has been renamed to delegates and now accepts an array instead of a single string.

- const model = await loadTensorflowModel(source, 'core-ml')
+ const model = await loadTensorflowModel(source, ['core-ml'])
- const model = useTensorflowModel(source, 'metal')
+ const model = useTensorflowModel(source, ['metal'])

The 'default' delegate is removed

To use CPU-only inference (previously 'default'), pass an empty array:

- const model = await loadTensorflowModel(source, 'default')
+ const model = await loadTensorflowModel(source, [])

model.delegate -> model.delegates

- console.log(model.delegate)   // 'core-ml'
+ console.log(model.delegates)  // ['core-ml']

3. Inputs/Outputs: TypedArray -> ArrayBuffer

run() and runSync() now accept and return ArrayBuffer[] instead of TypedArray[].

Preparing inputs

- const output = model.runSync([float32Array])
+ const output = model.runSync([float32Array.buffer])

If your TypedArray is a view into a larger buffer, slice it first:

const inputBuffer = typedArray.buffer.slice(
  typedArray.byteOffset,
  typedArray.byteOffset + typedArray.byteLength
)
const output = model.runSync([inputBuffer])

Reading outputs

  const outputs = model.runSync([input])
- const scores = outputs[0]  // was already a Float32Array
+ const scores = new Float32Array(outputs[0]!)  // wrap ArrayBuffer manually

Choose the appropriate TypedArray constructor based on your model's output tensor data type (e.g., Uint8Array, Int32Array, Float32Array).

4. VisionCamera v4 worklet compatibility

Because TfliteModel is now a Nitro HybridObject, VisionCamera v4's worklet runtime cannot access it directly. You need to box/unbox it:

import { NitroModules } from 'react-native-nitro-modules'

const model = await loadTensorflowModel(source, ['core-ml'])
const boxedModel = NitroModules.box(model)

// Inside a VisionCamera V4 frame processor worklet:
const unboxedModel = boxedModel.unbox()
const output = unboxedModel.runSync([inputBuffer])

This workaround will not be needed with VisionCamera v5.

5. Android gradle.properties

If you had custom property overrides using the Tflite_ prefix, rename them to NitroTflite_:

- Tflite_enableGpuDelegate=true
+ NitroTflite_enableGpuDelegate=true

6. Expanded TensorDataType

Tensor.dataType now includes additional types: string, bfloat16, int4, uint16, uint32, uint64, complex64, complex128, resource, variant, none.

The 'invalid' type has been removed. If you had exhaustive switch/case handling over dataType, update it accordingly.